+/*
+ * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
+ *
+ * This file is part of libFirm.
+ *
+ * This file may be distributed and/or modified under the terms of the
+ * GNU General Public License version 2 as published by the Free Software
+ * Foundation and appearing in the file LICENSE.GPL included in the
+ * packaging of this file.
+ *
+ * Licensees holding valid libFirm Professional Edition licenses may use
+ * this file in accordance with the libFirm Commercial License.
+ * Agreement provided with the Software.
+ *
+ * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
+ * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE.
+ */
+
+/**
+ * @file
+ * @brief Implements several optimizations for IA32.
+ * @author Matthias Braun, Christian Wuerdig
+ * @version $Id$
+ */
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#include "irnode.h"
#include "irprog_t.h"
#include "ircons.h"
+#include "irtools.h"
#include "firm_types.h"
#include "iredges.h"
#include "tv.h"
#include "irgmod.h"
+#include "irgwalk.h"
+#include "height.h"
+#include "irbitset.h"
+#include "irprintf.h"
+#include "error.h"
#include "../be_t.h"
#include "../beabi.h"
#include "../benode_t.h"
#include "../besched_t.h"
+#include "../bepeephole.h"
#include "ia32_new_nodes.h"
+#include "ia32_optimize.h"
#include "bearch_ia32_t.h"
-#include "gen_ia32_regalloc_if.h" /* the generated interface (register type and class defenitions) */
+#include "gen_ia32_regalloc_if.h"
+#include "ia32_common_transform.h"
+#include "ia32_transform.h"
+#include "ia32_dbg_stat.h"
+#include "ia32_util.h"
+#include "ia32_architecture.h"
-#undef is_NoMem
-#define is_NoMem(irn) (get_irn_op(irn) == op_NoMem)
+DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
-typedef int is_op_func_t(const ir_node *n);
+static ia32_code_gen_t *cg;
-static int be_is_NoReg(be_abi_irg_t *babi, const ir_node *irn) {
- if (be_abi_get_callee_save_irn(babi, &ia32_gp_regs[REG_XXX]) == irn ||
- be_abi_get_callee_save_irn(babi, &ia32_fp_regs[REG_XXXX]) == irn)
- {
- return 1;
- }
-
- return 0;
+static void copy_mark(const ir_node *old, ir_node *new)
+{
+ if (is_ia32_is_reload(old))
+ set_ia32_is_reload(new);
+ if (is_ia32_is_spill(old))
+ set_ia32_is_spill(new);
+ if (is_ia32_is_remat(old))
+ set_ia32_is_remat(new);
}
-
-
-/*************************************************
- * _____ _ _
- * / ____| | | | |
- * | | ___ _ __ ___| |_ __ _ _ __ | |_ ___
- * | | / _ \| '_ \/ __| __/ _` | '_ \| __/ __|
- * | |___| (_) | | | \__ \ || (_| | | | | |_\__ \
- * \_____\___/|_| |_|___/\__\__,_|_| |_|\__|___/
- *
- *************************************************/
+typedef enum produces_flag_t {
+ produces_no_flag,
+ produces_flag_zero,
+ produces_flag_carry
+} produces_flag_t;
/**
- * creates a unique ident by adding a number to a tag
+ * Return which usable flag the given node produces
*
- * @param tag the tag string, must contain a %d if a number
- * should be added
+ * @param node the node to check
+ * @param pn the projection number of the used result
*/
-static ident *unique_id(const char *tag)
+static produces_flag_t produces_test_flag(ir_node *node, int pn)
{
- static unsigned id = 0;
- char str[256];
-
- snprintf(str, sizeof(str), tag, ++id);
- return new_id_from_str(str);
-}
+ ir_node *count;
+ const ia32_immediate_attr_t *imm_attr;
+
+ if (!is_ia32_irn(node))
+ return produces_no_flag;
+
+ switch (get_ia32_irn_opcode(node)) {
+ case iro_ia32_Add:
+ case iro_ia32_Adc:
+ case iro_ia32_And:
+ case iro_ia32_Or:
+ case iro_ia32_Xor:
+ case iro_ia32_Sub:
+ case iro_ia32_Sbb:
+ case iro_ia32_Neg:
+ case iro_ia32_Inc:
+ case iro_ia32_Dec:
+ break;
+ case iro_ia32_ShlD:
+ case iro_ia32_ShrD:
+ assert(n_ia32_ShlD_count == n_ia32_ShrD_count);
+ count = get_irn_n(node, n_ia32_ShlD_count);
+ goto check_shift_amount;
+
+ case iro_ia32_Shl:
+ case iro_ia32_Shr:
+ case iro_ia32_Sar:
+ assert(n_ia32_Shl_count == n_ia32_Shr_count
+ && n_ia32_Shl_count == n_ia32_Sar_count);
+ count = get_irn_n(node, n_ia32_Shl_count);
+check_shift_amount:
+ /* when shift count is zero the flags are not affected, so we can only
+ * do this for constants != 0 */
+ if (!is_ia32_Immediate(count))
+ return produces_no_flag;
+
+ imm_attr = get_ia32_immediate_attr_const(count);
+ if (imm_attr->symconst != NULL)
+ return produces_no_flag;
+ if ((imm_attr->offset & 0x1f) == 0)
+ return produces_no_flag;
+ break;
+ case iro_ia32_Mul:
+ return pn == pn_ia32_Mul_res_high ?
+ produces_flag_carry : produces_no_flag;
-/**
- * Transforms a SymConst.
- *
- * @param mod the debug module
- * @param block the block the new node should belong to
- * @param node the ir SymConst node
- * @param mode mode of the SymConst
- * @return the created ia32 Const node
- */
-static ir_node *gen_SymConst(ia32_transform_env_t *env) {
- ir_node *cnst;
- dbg_info *dbg = env->dbg;
- ir_mode *mode = env->mode;
- ir_graph *irg = env->irg;
- ir_node *block = env->block;
-
- if (mode_is_float(mode)) {
- cnst = new_rd_ia32_fConst(dbg, irg, block, mode);
- }
- else {
- cnst = new_rd_ia32_Const(dbg, irg, block, mode);
+ default:
+ return produces_no_flag;
}
- set_ia32_Const_attr(cnst, env->irn);
- return cnst;
+
+ return pn == pn_ia32_res ?
+ produces_flag_zero : produces_no_flag;
}
/**
- * Get a primitive type for a mode.
+ * If the given node has not mode_T, creates a mode_T version (with a result Proj).
+ *
+ * @param node the node to change
+ *
+ * @return the new mode_T node (if the mode was changed) or node itself
*/
-static ir_type *get_prim_type(pmap *types, ir_mode *mode)
+static ir_node *turn_into_mode_t(ir_node *node)
{
- pmap_entry *e = pmap_find(types, mode);
- ir_type *res;
-
- if (! e) {
- char buf[64];
- snprintf(buf, sizeof(buf), "prim_type_%s", get_mode_name(mode));
- res = new_type_primitive(new_id_from_str(buf), mode);
- pmap_insert(types, mode, res);
- }
- else
- res = e->value;
- return res;
+ ir_node *block;
+ ir_node *res_proj;
+ ir_node *new_node;
+ const arch_register_t *reg;
+
+ if(get_irn_mode(node) == mode_T)
+ return node;
+
+ assert(get_irn_mode(node) == mode_Iu);
+
+ new_node = exact_copy(node);
+ set_irn_mode(new_node, mode_T);
+
+ block = get_nodes_block(new_node);
+ res_proj = new_r_Proj(current_ir_graph, block, new_node, mode_Iu,
+ pn_ia32_res);
+
+ reg = arch_get_irn_register(node);
+ arch_set_irn_register(res_proj, reg);
+
+ sched_add_before(node, new_node);
+ be_peephole_exchange(node, res_proj);
+ return new_node;
}
/**
- * Get an entity that is initialized with a tarval
+ * Replace Cmp(x, 0) by a Test(x, x)
*/
-static entity *get_entity_for_tv(ia32_code_gen_t *cg, ir_node *cnst)
+static void peephole_ia32_Cmp(ir_node *const node)
{
- tarval *tv = get_Const_tarval(cnst);
- pmap_entry *e = pmap_find(cg->tv_ent, tv);
- entity *res;
- ir_graph *rem;
-
- if (! e) {
- ir_mode *mode = get_irn_mode(cnst);
- ir_type *tp = get_Const_type(cnst);
- if (tp == firm_unknown_type)
- tp = get_prim_type(cg->types, mode);
-
- res = new_entity(get_glob_type(), unique_id("ia32FloatCnst_%u"), tp);
-
- set_entity_ld_ident(res, get_entity_ident(res));
- set_entity_visibility(res, visibility_local);
- set_entity_variability(res, variability_constant);
- set_entity_allocation(res, allocation_static);
-
- /* we create a new entity here: It's initialization must resist on the
- const code irg */
- rem = current_ir_graph;
- current_ir_graph = get_const_code_irg();
- set_atomic_ent_value(res, new_Const_type(tv, tp));
- current_ir_graph = rem;
- }
- else
- res = e->value;
- return res;
-}
+ ir_node *right;
+ ia32_immediate_attr_t const *imm;
+ dbg_info *dbgi;
+ ir_graph *irg;
+ ir_node *block;
+ ir_node *noreg;
+ ir_node *nomem;
+ ir_node *op;
+ ia32_attr_t const *attr;
+ int ins_permuted;
+ int cmp_unsigned;
+ ir_node *test;
+ arch_register_t const *reg;
+ ir_edge_t const *edge;
+ ir_edge_t const *tmp;
+
+ if (get_ia32_op_type(node) != ia32_Normal)
+ return;
-/**
- * Transforms a Const.
- *
- * @param mod the debug module
- * @param block the block the new node should belong to
- * @param node the ir Const node
- * @param mode mode of the Const
- * @return the created ia32 Const node
- */
-static ir_node *gen_Const(ia32_transform_env_t *env) {
- ir_node *cnst;
- symconst_symbol sym;
- ir_graph *irg = env->irg;
- ir_node *block = env->block;
- ir_node *node = env->irn;
- dbg_info *dbg = env->dbg;
- ir_mode *mode = env->mode;
-
- if (mode_is_float(mode)) {
- sym.entity_p = get_entity_for_tv(env->cg, node);
-
- cnst = new_rd_SymConst(dbg, irg, block, sym, symconst_addr_ent);
- env->irn = cnst;
- cnst = gen_SymConst(env);
- }
- else {
- cnst = new_rd_ia32_Const(dbg, irg, block, get_irn_mode(node));
- set_ia32_Const_attr(cnst, node);
+ right = get_irn_n(node, n_ia32_Cmp_right);
+ if (!is_ia32_Immediate(right))
+ return;
+
+ imm = get_ia32_immediate_attr_const(right);
+ if (imm->symconst != NULL || imm->offset != 0)
+ return;
+
+ dbgi = get_irn_dbg_info(node);
+ irg = current_ir_graph;
+ block = get_nodes_block(node);
+ noreg = ia32_new_NoReg_gp(cg);
+ nomem = get_irg_no_mem(irg);
+ op = get_irn_n(node, n_ia32_Cmp_left);
+ attr = get_irn_generic_attr(node);
+ ins_permuted = attr->data.ins_permuted;
+ cmp_unsigned = attr->data.cmp_unsigned;
+
+ if (is_ia32_Cmp(node)) {
+ test = new_rd_ia32_Test(dbgi, irg, block, noreg, noreg, nomem,
+ op, op, ins_permuted, cmp_unsigned);
+ } else {
+ test = new_rd_ia32_Test8Bit(dbgi, irg, block, noreg, noreg, nomem,
+ op, op, ins_permuted, cmp_unsigned);
}
- return cnst;
-}
+ set_ia32_ls_mode(test, get_ia32_ls_mode(node));
+ reg = arch_get_irn_register(node);
+ arch_set_irn_register(test, reg);
+ foreach_out_edge_safe(node, edge, tmp) {
+ ir_node *const user = get_edge_src_irn(edge);
+
+ if (is_Proj(user))
+ exchange(user, test);
+ }
+
+ sched_add_before(node, test);
+ copy_mark(node, test);
+ be_peephole_exchange(node, test);
+}
/**
- * Transforms (all) Const's into ia32_Const and places them in the
- * block where they are used (or in the cfg-pred Block in case of Phi's).
- * Additionally all reference nodes are changed into mode_Is nodes.
+ * Peephole optimization for Test instructions.
+ * We can remove the Test, if a zero flags was produced which is still
+ * live.
*/
-void ia32_place_consts_set_modes(ir_node *irn, void *env) {
- ia32_code_gen_t *cg = env;
- ia32_transform_env_t tenv;
- ir_mode *mode;
- ir_node *pred, *cnst;
- int i;
- opcode opc;
-
- if (is_Block(irn))
+static void peephole_ia32_Test(ir_node *node)
+{
+ ir_node *left = get_irn_n(node, n_ia32_Test_left);
+ ir_node *right = get_irn_n(node, n_ia32_Test_right);
+ ir_node *flags_proj;
+ ir_node *block;
+ ir_mode *flags_mode;
+ int pn = pn_ia32_res;
+ ir_node *schedpoint;
+ const ir_edge_t *edge;
+
+ assert(n_ia32_Test_left == n_ia32_Test8Bit_left
+ && n_ia32_Test_right == n_ia32_Test8Bit_right);
+
+ /* we need a test for 0 */
+ if(left != right)
return;
- mode = get_irn_mode(irn);
+ block = get_nodes_block(node);
+ if(get_nodes_block(left) != block)
+ return;
- /* transform all reference nodes into mode_Is nodes */
- if (mode_is_reference(mode)) {
- mode = mode_Is;
- set_irn_mode(irn, mode);
+ if(is_Proj(left)) {
+ pn = get_Proj_proj(left);
+ left = get_Proj_pred(left);
}
- tenv.block = get_nodes_block(irn);
- tenv.cg = cg;
- tenv.irg = cg->irg;
- tenv.mod = cg->mod;
-
- /* Loop over all predecessors and check for Sym/Const nodes */
- for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
- pred = get_irn_n(irn, i);
- cnst = NULL;
- opc = get_irn_opcode(pred);
- tenv.irn = pred;
- tenv.mode = get_irn_mode(pred);
- tenv.dbg = get_irn_dbg_info(pred);
-
- /* If it's a Phi, then we need to create the */
- /* new Const in it's predecessor block */
- if (is_Phi(irn)) {
- tenv.block = get_Block_cfgpred_block(get_nodes_block(irn), i);
- }
+ /* happens rarely, but if it does code will panic' */
+ if (is_ia32_Unknown_GP(left))
+ return;
- /* put the const into the block where the original const was */
- if (! cg->opt.placecnst) {
- tenv.block = get_nodes_block(pred);
- }
+ /* walk schedule up and abort when we find left or some other node destroys
+ the flags */
+ schedpoint = node;
+ for (;;) {
+ schedpoint = sched_prev(schedpoint);
+ if (schedpoint == left)
+ break;
+ if (arch_irn_is(cg->arch_env, schedpoint, modify_flags))
+ return;
+ if (schedpoint == block)
+ panic("couldn't find left");
+ }
- switch (opc) {
- case iro_Const:
- cnst = gen_Const(&tenv);
- break;
- case iro_SymConst:
- cnst = gen_SymConst(&tenv);
- break;
- default:
- break;
- }
+ /* make sure only Lg/Eq tests are used */
+ foreach_out_edge(node, edge) {
+ ir_node *user = get_edge_src_irn(edge);
+ int pnc = get_ia32_condcode(user);
- /* if we found a const, then set it */
- if (cnst) {
- set_irn_n(irn, i, cnst);
+ if(pnc != pn_Cmp_Eq && pnc != pn_Cmp_Lg) {
+ return;
}
}
-}
+ switch (produces_test_flag(left, pn)) {
+ case produces_flag_zero:
+ break;
+ case produces_flag_carry:
+ foreach_out_edge(node, edge) {
+ ir_node *user = get_edge_src_irn(edge);
+ int pnc = get_ia32_condcode(user);
-/********************************************************************************************************
- * _____ _ _ ____ _ _ _ _ _
- * | __ \ | | | | / __ \ | | (_) (_) | | (_)
- * | |__) |__ ___ _ __ | |__ ___ | | ___ | | | |_ __ | |_ _ _ __ ___ _ ______ _| |_ _ ___ _ __
- * | ___/ _ \/ _ \ '_ \| '_ \ / _ \| |/ _ \ | | | | '_ \| __| | '_ ` _ \| |_ / _` | __| |/ _ \| '_ \
- * | | | __/ __/ |_) | | | | (_) | | __/ | |__| | |_) | |_| | | | | | | |/ / (_| | |_| | (_) | | | |
- * |_| \___|\___| .__/|_| |_|\___/|_|\___| \____/| .__/ \__|_|_| |_| |_|_/___\__,_|\__|_|\___/|_| |_|
- * | | | |
- * |_| |_|
- ********************************************************************************************************/
+ switch (pnc) {
+ case pn_Cmp_Eq: pnc = pn_Cmp_Ge | ia32_pn_Cmp_unsigned; break;
+ case pn_Cmp_Lg: pnc = pn_Cmp_Lt | ia32_pn_Cmp_unsigned; break;
+ default: panic("unexpected pn");
+ }
+ set_ia32_condcode(user, pnc);
+ }
+ break;
-/**
- * NOTE: THESE PEEPHOLE OPTIMIZATIONS MUST BE CALLED AFTER SCHEDULING AND REGISTER ALLOCATION.
- */
+ default:
+ return;
+ }
+
+ left = turn_into_mode_t(left);
+
+ flags_mode = ia32_reg_classes[CLASS_ia32_flags].mode;
+ flags_proj = new_r_Proj(current_ir_graph, block, left, flags_mode,
+ pn_ia32_flags);
+ arch_set_irn_register(flags_proj, &ia32_flags_regs[REG_EFLAGS]);
-static int ia32_cnst_compare(ir_node *n1, ir_node *n2) {
- return get_ia32_id_cnst(n1) == get_ia32_id_cnst(n2);
+ assert(get_irn_mode(node) != mode_T);
+
+ be_peephole_exchange(node, flags_proj);
}
/**
- * Checks for potential CJmp/CJmpAM optimization candidates.
+ * AMD Athlon works faster when RET is not destination of
+ * conditional jump or directly preceded by other jump instruction.
+ * Can be avoided by placing a Rep prefix before the return.
*/
-static ir_node *ia32_determine_cjmp_cand(ir_node *irn, is_op_func_t *is_op_func) {
- ir_node *cand = NULL;
- ir_node *prev = sched_prev(irn);
-
- if (is_Block(prev)) {
- if (get_Block_n_cfgpreds(prev) == 1)
- prev = get_Block_cfgpred(prev, 0);
- else
- prev = NULL;
- }
+static void peephole_ia32_Return(ir_node *node) {
+ ir_node *block, *irn;
- /* The predecessor must be a ProjX. */
- if (prev && is_Proj(prev) && get_irn_mode(prev) == mode_X) {
- prev = get_Proj_pred(prev);
+ if (!ia32_cg_config.use_pad_return)
+ return;
- if (is_op_func(prev))
- cand = prev;
+ block = get_nodes_block(node);
+
+ /* check if this return is the first on the block */
+ sched_foreach_reverse_from(node, irn) {
+ switch (get_irn_opcode(irn)) {
+ case beo_Return:
+ /* the return node itself, ignore */
+ continue;
+ case iro_Start:
+ case beo_RegParams:
+ case beo_Barrier:
+ /* ignore no code generated */
+ continue;
+ case beo_IncSP:
+ /* arg, IncSP 0 nodes might occur, ignore these */
+ if (be_get_IncSP_offset(irn) == 0)
+ continue;
+ return;
+ case iro_Phi:
+ continue;
+ default:
+ return;
+ }
}
- return cand;
+ /* ensure, that the 3 byte return is generated */
+ be_Return_set_emit_pop(node, 1);
}
-static int is_TestJmp_cand(const ir_node *irn) {
- return is_ia32_TestJmp(irn) || is_ia32_And(irn);
-}
+/* only optimize up to 48 stores behind IncSPs */
+#define MAXPUSH_OPTIMIZE 48
/**
- * Checks if two consecutive arguments of cand matches
- * the two arguments of irn (TestJmp).
+ * Tries to create Push's from IncSP, Store combinations.
+ * The Stores are replaced by Push's, the IncSP is modified
+ * (possibly into IncSP 0, but not removed).
*/
-static int is_TestJmp_replacement(ir_node *cand, ir_node *irn) {
- ir_node *in1 = get_irn_n(irn, 0);
- ir_node *in2 = get_irn_n(irn, 1);
- int i, n = get_irn_arity(cand);
- int same_args = 0;
-
- for (i = 0; i < n - 1; i++) {
- if (get_irn_n(cand, i) == in1 &&
- get_irn_n(cand, i + 1) == in2)
- {
- same_args = 1;
+static void peephole_IncSP_Store_to_push(ir_node *irn)
+{
+ int i;
+ int maxslot;
+ int inc_ofs;
+ ir_node *node;
+ ir_node *stores[MAXPUSH_OPTIMIZE];
+ ir_node *block;
+ ir_graph *irg;
+ ir_node *curr_sp;
+ ir_mode *spmode;
+ ir_node *first_push = NULL;
+ ir_edge_t const *edge;
+ ir_edge_t const *next;
+
+ memset(stores, 0, sizeof(stores));
+
+ assert(be_is_IncSP(irn));
+
+ inc_ofs = be_get_IncSP_offset(irn);
+ if (inc_ofs < 4)
+ return;
+
+ /*
+ * We first walk the schedule after the IncSP node as long as we find
+ * suitable Stores that could be transformed to a Push.
+ * We save them into the stores array which is sorted by the frame offset/4
+ * attached to the node
+ */
+ maxslot = -1;
+ for (node = sched_next(irn); !sched_is_end(node); node = sched_next(node)) {
+ ir_node *mem;
+ int offset;
+ int storeslot;
+
+ /* it has to be a Store */
+ if (!is_ia32_Store(node))
break;
- }
+
+ /* it has to use our sp value */
+ if (get_irn_n(node, n_ia32_base) != irn)
+ continue;
+ /* Store has to be attached to NoMem */
+ mem = get_irn_n(node, n_ia32_mem);
+ if (!is_NoMem(mem))
+ continue;
+
+ /* unfortunately we can't support the full AMs possible for push at the
+ * moment. TODO: fix this */
+ if (!is_ia32_NoReg_GP(get_irn_n(node, n_ia32_index)))
+ break;
+
+ offset = get_ia32_am_offs_int(node);
+ /* we should NEVER access uninitialized stack BELOW the current SP */
+ assert(offset >= 0);
+
+ /* storing at half-slots is bad */
+ if ((offset & 3) != 0)
+ break;
+
+ if (inc_ofs - 4 < offset || offset >= MAXPUSH_OPTIMIZE * 4)
+ continue;
+ storeslot = offset >> 2;
+
+ /* storing into the same slot twice is bad (and shouldn't happen...) */
+ if (stores[storeslot] != NULL)
+ break;
+
+ stores[storeslot] = node;
+ if (storeslot > maxslot)
+ maxslot = storeslot;
}
- if (same_args)
- return ia32_cnst_compare(cand, irn);
+ curr_sp = irn;
- return 0;
-}
+ for (i = -1; i < maxslot; ++i) {
+ if (stores[i + 1] == NULL)
+ break;
+ }
-/**
- * Tries to replace a TestJmp by a CJmp or CJmpAM (in case of And)
- */
-static void ia32_optimize_TestJmp(ir_node *irn, ia32_code_gen_t *cg) {
- ir_node *cand = ia32_determine_cjmp_cand(irn, is_TestJmp_cand);
- int replace = 0;
+ /* walk through the Stores and create Pushs for them */
+ block = get_nodes_block(irn);
+ spmode = get_irn_mode(irn);
+ irg = cg->irg;
+ for (; i >= 0; --i) {
+ const arch_register_t *spreg;
+ ir_node *push;
+ ir_node *val, *mem, *mem_proj;
+ ir_node *store = stores[i];
+ ir_node *noreg = ia32_new_NoReg_gp(cg);
- /* we found a possible candidate */
- replace = cand ? is_TestJmp_replacement(cand, irn) : 0;
+ val = get_irn_n(store, n_ia32_unary_op);
+ mem = get_irn_n(store, n_ia32_mem);
+ spreg = arch_get_irn_register(curr_sp);
- if (replace) {
- DBG((cg->mod, LEVEL_1, "replacing %+F by ", irn));
+ push = new_rd_ia32_Push(get_irn_dbg_info(store), irg, block, noreg, noreg, mem, val, curr_sp);
+ copy_mark(store, push);
- if (is_ia32_And(cand))
- set_irn_op(irn, op_ia32_CJmpAM);
- else
- set_irn_op(irn, op_ia32_CJmp);
+ if (first_push == NULL)
+ first_push = push;
- DB((cg->mod, LEVEL_1, "%+F\n", irn));
+ sched_add_after(curr_sp, push);
+
+ /* create stackpointer Proj */
+ curr_sp = new_r_Proj(irg, block, push, spmode, pn_ia32_Push_stack);
+ arch_set_irn_register(curr_sp, spreg);
+
+ /* create memory Proj */
+ mem_proj = new_r_Proj(irg, block, push, mode_M, pn_ia32_Push_M);
+
+ /* use the memproj now */
+ be_peephole_exchange(store, mem_proj);
+
+ inc_ofs -= 4;
}
-}
-static int is_CondJmp_cand(const ir_node *irn) {
- return is_ia32_CondJmp(irn) || is_ia32_Sub(irn);
+ foreach_out_edge_safe(irn, edge, next) {
+ ir_node *const src = get_edge_src_irn(edge);
+ int const pos = get_edge_src_pos(edge);
+
+ if (src == first_push)
+ continue;
+
+ set_irn_n(src, pos, curr_sp);
+ }
+
+ be_set_IncSP_offset(irn, inc_ofs);
}
-/**
- * Checks if the arguments of cand are the same of irn.
- */
-static int is_CondJmp_replacement(ir_node *cand, ir_node *irn) {
- int i, n = get_irn_arity(cand);
- int same_args = 0;
+#if 0
+static void peephole_store_incsp(ir_node *store)
+{
+ dbg_info *dbgi;
+ ir_node *node;
+ ir_node *block;
+ ir_node *noref;
+ ir_node *mem;
+ ir_node *push;
+ ir_node *val;
+ ir_node *am_base = get_irn_n(store, n_ia32_Store_base);
+ if (!be_is_IncSP(am_base)
+ || get_nodes_block(am_base) != get_nodes_block(store))
+ return;
+ mem = get_irn_n(store, n_ia32_Store_mem);
+ if (!is_ia32_NoReg_GP(get_irn_n(store, n_ia32_Store_index))
+ || !is_NoMem(mem))
+ return;
+
+ int incsp_offset = be_get_IncSP_offset(am_base);
+ if (incsp_offset <= 0)
+ return;
+
+ /* we have to be at offset 0 */
+ int my_offset = get_ia32_am_offs_int(store);
+ if (my_offset != 0) {
+ /* TODO here: find out wether there is a store with offset 0 before
+ * us and wether we can move it down to our place */
+ return;
+ }
+ ir_mode *ls_mode = get_ia32_ls_mode(store);
+ int my_store_size = get_mode_size_bytes(ls_mode);
+
+ if (my_offset + my_store_size > incsp_offset)
+ return;
- for (i = 0; i < n; i++) {
- if (get_irn_n(cand, i) == get_irn_n(irn, i)) {
- same_args = 1;
+ /* correctness checking:
+ - noone else must write to that stackslot
+ (because after translation incsp won't allocate it anymore)
+ */
+ sched_foreach_reverse_from(store, node) {
+ int i, arity;
+
+ if (node == am_base)
break;
+
+ /* make sure noone else can use the space on the stack */
+ arity = get_irn_arity(node);
+ for (i = 0; i < arity; ++i) {
+ ir_node *pred = get_irn_n(node, i);
+ if (pred != am_base)
+ continue;
+
+ if (i == n_ia32_base &&
+ (get_ia32_op_type(node) == ia32_AddrModeS
+ || get_ia32_op_type(node) == ia32_AddrModeD)) {
+ int node_offset = get_ia32_am_offs_int(node);
+ ir_mode *node_ls_mode = get_ia32_ls_mode(node);
+ int node_size = get_mode_size_bytes(node);
+ /* overlapping with our position? abort */
+ if (node_offset < my_offset + my_store_size
+ && node_offset + node_size >= my_offset)
+ return;
+ /* otherwise it's fine */
+ continue;
+ }
+
+ /* strange use of esp: abort */
+ return;
}
}
- if (same_args)
- return ia32_cnst_compare(cand, irn);
+ /* all ok, change to push */
+ dbgi = get_irn_dbg_info(store);
+ block = get_nodes_block(store);
+ noreg = ia32_new_NoReg_gp(cg);
+ val = get_ia32_
+
+ push = new_rd_ia32_Push(dbgi, irg, block, noreg, noreg, mem,
- return 0;
+ create_push(dbgi, current_ir_graph, block, am_base, store);
+}
+#endif
+
+/**
+ * Return true if a mode can be stored in the GP register set
+ */
+static INLINE int mode_needs_gp_reg(ir_mode *mode) {
+ if (mode == mode_fpcw)
+ return 0;
+ if (get_mode_size_bits(mode) > 32)
+ return 0;
+ return mode_is_int(mode) || mode_is_reference(mode) || mode == mode_b;
}
/**
- * Tries to replace a CondJmp by a CJmpAM
+ * Tries to create Pops from Load, IncSP combinations.
+ * The Loads are replaced by Pops, the IncSP is modified
+ * (possibly into IncSP 0, but not removed).
*/
-static void ia32_optimize_CondJmp(ir_node *irn, ia32_code_gen_t *cg) {
- ir_node *cand = ia32_determine_cjmp_cand(irn, is_CondJmp_cand);
- int replace = 0;
+static void peephole_Load_IncSP_to_pop(ir_node *irn)
+{
+ const arch_register_t *esp = &ia32_gp_regs[REG_ESP];
+ int i, maxslot, inc_ofs, ofs;
+ ir_node *node, *pred_sp, *block;
+ ir_node *loads[MAXPUSH_OPTIMIZE];
+ ir_graph *irg;
+ unsigned regmask = 0;
+ unsigned copymask = ~0;
+
+ memset(loads, 0, sizeof(loads));
+ assert(be_is_IncSP(irn));
+
+ inc_ofs = -be_get_IncSP_offset(irn);
+ if (inc_ofs < 4)
+ return;
- /* we found a possible candidate */
- replace = cand ? is_CondJmp_replacement(cand, irn) : 0;
+ /*
+ * We first walk the schedule before the IncSP node as long as we find
+ * suitable Loads that could be transformed to a Pop.
+ * We save them into the stores array which is sorted by the frame offset/4
+ * attached to the node
+ */
+ maxslot = -1;
+ pred_sp = be_get_IncSP_pred(irn);
+ for (node = sched_prev(irn); !sched_is_end(node); node = sched_prev(node)) {
+ int offset;
+ int loadslot;
+ const arch_register_t *sreg, *dreg;
+
+ /* it has to be a Load */
+ if (!is_ia32_Load(node)) {
+ if (be_is_Copy(node)) {
+ if (!mode_needs_gp_reg(get_irn_mode(node))) {
+ /* not a GP copy, ignore */
+ continue;
+ }
+ dreg = arch_get_irn_register(node);
+ sreg = arch_get_irn_register(be_get_Copy_op(node));
+ if (regmask & copymask & (1 << sreg->index)) {
+ break;
+ }
+ if (regmask & copymask & (1 << dreg->index)) {
+ break;
+ }
+ /* we CAN skip Copies if neither the destination nor the source
+ * is not in our regmask, ie none of our future Pop will overwrite it */
+ regmask |= (1 << dreg->index) | (1 << sreg->index);
+ copymask &= ~((1 << dreg->index) | (1 << sreg->index));
+ continue;
+ }
+ break;
+ }
+
+ /* we can handle only GP loads */
+ if (!mode_needs_gp_reg(get_ia32_ls_mode(node)))
+ continue;
- if (replace) {
- DBG((cg->mod, LEVEL_1, "replacing %+F by ", irn));
+ /* it has to use our predecessor sp value */
+ if (get_irn_n(node, n_ia32_base) != pred_sp) {
+ /* it would be ok if this load does not use a Pop result,
+ * but we do not check this */
+ break;
+ }
- set_irn_op(irn, op_ia32_CJmp);
+ /* should have NO index */
+ if (!is_ia32_NoReg_GP(get_irn_n(node, n_ia32_index)))
+ break;
+
+ offset = get_ia32_am_offs_int(node);
+ /* we should NEVER access uninitialized stack BELOW the current SP */
+ assert(offset >= 0);
+
+ /* storing at half-slots is bad */
+ if ((offset & 3) != 0)
+ break;
+
+ if (offset < 0 || offset >= MAXPUSH_OPTIMIZE * 4)
+ continue;
+ /* ignore those outside the possible windows */
+ if (offset > inc_ofs - 4)
+ continue;
+ loadslot = offset >> 2;
+
+ /* loading from the same slot twice is bad (and shouldn't happen...) */
+ if (loads[loadslot] != NULL)
+ break;
+
+ dreg = arch_get_irn_register(node);
+ if (regmask & (1 << dreg->index)) {
+ /* this register is already used */
+ break;
+ }
+ regmask |= 1 << dreg->index;
- DB((cg->mod, LEVEL_1, "%+F\n", irn));
+ loads[loadslot] = node;
+ if (loadslot > maxslot)
+ maxslot = loadslot;
}
-}
-/**
- * Performs Peephole Optimizations
- */
-void ia32_peephole_optimization(ir_node *irn, void *env) {
- if (is_ia32_TestJmp(irn)) {
- ia32_optimize_TestJmp(irn, env);
+ if (maxslot < 0)
+ return;
+
+ /* find the first slot */
+ for (i = maxslot; i >= 0; --i) {
+ ir_node *load = loads[i];
+
+ if (load == NULL)
+ break;
}
- else if (is_ia32_CondJmp(irn)) {
- ia32_optimize_CondJmp(irn, env);
+
+ ofs = inc_ofs - (maxslot + 1) * 4;
+ inc_ofs = (i+1) * 4;
+
+ /* create a new IncSP if needed */
+ block = get_nodes_block(irn);
+ irg = cg->irg;
+ if (inc_ofs > 0) {
+ pred_sp = be_new_IncSP(esp, irg, block, pred_sp, -inc_ofs, be_get_IncSP_align(irn));
+ sched_add_before(irn, pred_sp);
}
-}
+ /* walk through the Loads and create Pops for them */
+ for (++i; i <= maxslot; ++i) {
+ ir_node *load = loads[i];
+ ir_node *mem, *pop;
+ const ir_edge_t *edge, *tmp;
+ const arch_register_t *reg;
+ mem = get_irn_n(load, n_ia32_mem);
+ reg = arch_get_irn_register(load);
-/******************************************************************
- * _ _ __ __ _
- * /\ | | | | | \/ | | |
- * / \ __| | __| |_ __ ___ ___ ___| \ / | ___ __| | ___
- * / /\ \ / _` |/ _` | '__/ _ \/ __/ __| |\/| |/ _ \ / _` |/ _ \
- * / ____ \ (_| | (_| | | | __/\__ \__ \ | | | (_) | (_| | __/
- * /_/ \_\__,_|\__,_|_| \___||___/___/_| |_|\___/ \__,_|\___|
- *
- ******************************************************************/
+ pop = new_rd_ia32_Pop(get_irn_dbg_info(load), irg, block, mem, pred_sp);
+ arch_set_irn_register(pop, reg);
-static int node_is_comm(const ir_node *irn) {
- return is_ia32_irn(irn) ? is_ia32_commutative(irn) : 0;
-}
+ copy_mark(load, pop);
-static int ia32_get_irn_n_edges(const ir_node *irn) {
- const ir_edge_t *edge;
- int cnt = 0;
+ /* create stackpointer Proj */
+ pred_sp = new_r_Proj(irg, block, pop, mode_Iu, pn_ia32_Pop_stack);
+ arch_set_irn_register(pred_sp, esp);
+
+ sched_add_before(irn, pop);
+
+ /* rewire now */
+ foreach_out_edge_safe(load, edge, tmp) {
+ ir_node *proj = get_edge_src_irn(edge);
+
+ set_Proj_pred(proj, pop);
+ }
- foreach_out_edge(irn, edge) {
- cnt++;
+ /* we can remove the Load now */
+ sched_remove(load);
+ kill_node(load);
}
- return cnt;
+ be_set_IncSP_offset(irn, -ofs);
+ be_set_IncSP_pred(irn, pred_sp);
}
+
/**
- * Returns the first mode_M Proj connected to irn.
+ * Find a free GP register if possible, else return NULL.
*/
-static ir_node *get_mem_proj(const ir_node *irn) {
- const ir_edge_t *edge;
- ir_node *src;
-
- assert(get_irn_mode(irn) == mode_T && "expected mode_T node");
-
- foreach_out_edge(irn, edge) {
- src = get_edge_src_irn(edge);
+static const arch_register_t *get_free_gp_reg(void)
+{
+ int i;
- assert(is_Proj(src) && "Proj expected");
+ for(i = 0; i < N_ia32_gp_REGS; ++i) {
+ const arch_register_t *reg = &ia32_gp_regs[i];
+ if(arch_register_type_is(reg, ignore))
+ continue;
- if (get_irn_mode(src) == mode_M)
- return src;
+ if(be_peephole_get_value(CLASS_ia32_gp, i) == NULL)
+ return &ia32_gp_regs[i];
}
return NULL;
}
/**
- * Returns the first Proj with mode != mode_M connected to irn.
+ * Creates a Pop instruction before the given schedule point.
+ *
+ * @param dbgi debug info
+ * @param irg the graph
+ * @param block the block
+ * @param stack the previous stack value
+ * @param schedpoint the new node is added before this node
+ * @param reg the register to pop
+ *
+ * @return the new stack value
*/
-static ir_node *get_res_proj(const ir_node *irn) {
- const ir_edge_t *edge;
- ir_node *src;
+static ir_node *create_pop(dbg_info *dbgi, ir_graph *irg, ir_node *block,
+ ir_node *stack, ir_node *schedpoint,
+ const arch_register_t *reg)
+{
+ const arch_register_t *esp = &ia32_gp_regs[REG_ESP];
+ ir_node *pop;
+ ir_node *keep;
+ ir_node *val;
+ ir_node *in[1];
- assert(get_irn_mode(irn) == mode_T && "expected mode_T node");
+ pop = new_rd_ia32_Pop(dbgi, irg, block, new_NoMem(), stack);
- foreach_out_edge(irn, edge) {
- src = get_edge_src_irn(edge);
+ stack = new_r_Proj(irg, block, pop, mode_Iu, pn_ia32_Pop_stack);
+ arch_set_irn_register(stack, esp);
+ val = new_r_Proj(irg, block, pop, mode_Iu, pn_ia32_Pop_res);
+ arch_set_irn_register(val, reg);
- assert(is_Proj(src) && "Proj expected");
+ sched_add_before(schedpoint, pop);
- if (get_irn_mode(src) != mode_M)
- return src;
- }
+ in[0] = val;
+ keep = be_new_Keep(&ia32_reg_classes[CLASS_ia32_gp], irg, block, 1, in);
+ sched_add_before(schedpoint, keep);
- return NULL;
+ return stack;
}
/**
- * Determines if pred is a Proj and if is_op_func returns true for it's predecessor.
+ * Creates a Push instruction before the given schedule point.
+ *
+ * @param dbgi debug info
+ * @param irg the graph
+ * @param block the block
+ * @param stack the previous stack value
+ * @param schedpoint the new node is added before this node
+ * @param reg the register to pop
*
- * @param pred The node to be checked
- * @param is_op_func The check-function
- * @return 1 if conditions are fulfilled, 0 otherwise
+ * @return the new stack value
*/
-static int pred_is_specific_node(const ir_node *pred, is_op_func_t *is_op_func) {
- if (is_Proj(pred) && is_op_func(get_Proj_pred(pred))) {
- return 1;
- }
+static ir_node *create_push(dbg_info *dbgi, ir_graph *irg, ir_node *block,
+ ir_node *stack, ir_node *schedpoint)
+{
+ const arch_register_t *esp = &ia32_gp_regs[REG_ESP];
+
+ ir_node *val = ia32_new_Unknown_gp(cg);
+ ir_node *noreg = ia32_new_NoReg_gp(cg);
+ ir_node *nomem = get_irg_no_mem(irg);
+ ir_node *push = new_rd_ia32_Push(dbgi, irg, block, noreg, noreg, nomem, val, stack);
+ sched_add_before(schedpoint, push);
- return 0;
+ stack = new_r_Proj(irg, block, push, mode_Iu, pn_ia32_Push_stack);
+ arch_set_irn_register(stack, esp);
+
+ return stack;
}
/**
- * Determines if pred is a Proj and if is_op_func returns true for it's predecessor
- * and if the predecessor is in block bl.
- *
- * @param bl The block
- * @param pred The node to be checked
- * @param is_op_func The check-function
- * @return 1 if conditions are fulfilled, 0 otherwise
+ * Optimize an IncSp by replacing it with Push/Pop.
*/
-static int pred_is_specific_nodeblock(const ir_node *bl, const ir_node *pred,
- int (*is_op_func)(const ir_node *n))
+static void peephole_be_IncSP(ir_node *node)
{
- if (is_Proj(pred)) {
- pred = get_Proj_pred(pred);
- if ((bl == get_nodes_block(pred)) && is_op_func(pred)) {
- return 1;
- }
- }
+ const arch_register_t *esp = &ia32_gp_regs[REG_ESP];
+ const arch_register_t *reg;
+ ir_graph *irg = current_ir_graph;
+ dbg_info *dbgi;
+ ir_node *block;
+ ir_node *stack;
+ int offset;
- return 0;
-}
+ /* first optimize incsp->incsp combinations */
+ node = be_peephole_IncSP_IncSP(node);
+ /* transform IncSP->Store combinations to Push where possible */
+ peephole_IncSP_Store_to_push(node);
+ /* transform Load->IncSP combinations to Pop where possible */
+ peephole_Load_IncSP_to_pop(node);
-/**
- * Checks if irn is a candidate for address calculation or address mode.
- *
- * address calculation (AC):
- * - none of the operand must be a Load within the same block OR
- * - all Loads must have more than one user OR
- * - the irn has a frame entity (it's a former FrameAddr)
- *
- * address mode (AM):
- * - at least one operand has to be a Load within the same block AND
- * - the load must not have other users than the irn AND
- * - the irn must not have a frame entity set
- *
- * @param block The block the Loads must/not be in
- * @param irn The irn to check
- * @param check_addr 1 if to check for address calculation, 0 otherwise
- * return 1 if irn is a candidate for AC or AM, 0 otherwise
- */
-static int is_candidate(const ir_node *block, const ir_node *irn, int check_addr) {
- ir_node *in;
- int n, is_cand = check_addr;
+ if (arch_get_irn_register(node) != esp)
+ return;
- in = get_irn_n(irn, 2);
+ /* replace IncSP -4 by Pop freereg when possible */
+ offset = be_get_IncSP_offset(node);
+ if ((offset != -8 || ia32_cg_config.use_add_esp_8) &&
+ (offset != -4 || ia32_cg_config.use_add_esp_4) &&
+ (offset != +4 || ia32_cg_config.use_sub_esp_4) &&
+ (offset != +8 || ia32_cg_config.use_sub_esp_8))
+ return;
- if (pred_is_specific_nodeblock(block, in, is_ia32_Ld)) {
- n = ia32_get_irn_n_edges(in);
- is_cand = check_addr ? (n == 1 ? 0 : is_cand) : (n == 1 ? 1 : is_cand);
- }
+ if (offset < 0) {
+ /* we need a free register for pop */
+ reg = get_free_gp_reg();
+ if (reg == NULL)
+ return;
- in = get_irn_n(irn, 3);
+ dbgi = get_irn_dbg_info(node);
+ block = get_nodes_block(node);
+ stack = be_get_IncSP_pred(node);
- if (pred_is_specific_nodeblock(block, in, is_ia32_Ld)) {
- n = ia32_get_irn_n_edges(in);
- is_cand = check_addr ? (n == 1 ? 0 : is_cand) : (n == 1 ? 1 : is_cand);
- }
+ stack = create_pop(dbgi, irg, block, stack, node, reg);
- is_cand = get_ia32_frame_ent(irn) ? (check_addr ? 1 : 0) : is_cand;
+ if (offset == -8) {
+ stack = create_pop(dbgi, irg, block, stack, node, reg);
+ }
+ } else {
+ dbgi = get_irn_dbg_info(node);
+ block = get_nodes_block(node);
+ stack = be_get_IncSP_pred(node);
+ stack = create_push(dbgi, irg, block, stack, node);
+
+ if (offset == +8) {
+ stack = create_push(dbgi, irg, block, stack, node);
+ }
+ }
- return is_cand;
+ be_peephole_exchange(node, stack);
}
/**
- * Compares the base and index addr and the load/store entities
- * and returns 1 if they are equal.
+ * Peephole optimisation for ia32_Const's
*/
-static int load_store_addr_is_equal(const ir_node *load, const ir_node *store,
- const ir_node *addr_b, const ir_node *addr_i)
+static void peephole_ia32_Const(ir_node *node)
{
- int is_equal = (addr_b == get_irn_n(load, 0)) && (addr_i == get_irn_n(load, 1));
- entity *lent = get_ia32_frame_ent(load);
- entity *sent = get_ia32_frame_ent(store);
-
- /* are both entities set and equal? */
- is_equal = lent && sent && (lent == sent);
+ const ia32_immediate_attr_t *attr = get_ia32_immediate_attr_const(node);
+ const arch_register_t *reg;
+ ir_graph *irg = current_ir_graph;
+ ir_node *block;
+ dbg_info *dbgi;
+ ir_node *produceval;
+ ir_node *xor;
+ ir_node *noreg;
+
+ /* try to transform a mov 0, reg to xor reg reg */
+ if (attr->offset != 0 || attr->symconst != NULL)
+ return;
+ if (ia32_cg_config.use_mov_0)
+ return;
+ /* xor destroys the flags, so no-one must be using them */
+ if (be_peephole_get_value(CLASS_ia32_flags, REG_EFLAGS) != NULL)
+ return;
- /* are the load and the store of the same mode? */
- is_equal = get_ia32_ls_mode(load) == get_ia32_ls_mode(store);
+ reg = arch_get_irn_register(node);
+ assert(be_peephole_get_reg_value(reg) == NULL);
- return is_equal;
-}
+ /* create xor(produceval, produceval) */
+ block = get_nodes_block(node);
+ dbgi = get_irn_dbg_info(node);
+ produceval = new_rd_ia32_ProduceVal(dbgi, irg, block);
+ arch_set_irn_register(produceval, reg);
+ noreg = ia32_new_NoReg_gp(cg);
+ xor = new_rd_ia32_Xor(dbgi, irg, block, noreg, noreg, new_NoMem(),
+ produceval, produceval);
+ arch_set_irn_register(xor, reg);
+ sched_add_before(node, produceval);
+ sched_add_before(node, xor);
-/**
- * Folds Add or Sub to LEA if possible
- */
-static ir_node *fold_addr(be_abi_irg_t *babi, ir_node *irn, firm_dbg_module_t *mod, ir_node *noreg) {
- ir_graph *irg = get_irn_irg(irn);
- dbg_info *dbg = get_irn_dbg_info(irn);
- ir_node *block = get_nodes_block(irn);
- ir_node *res = irn;
- char *offs = NULL;
- const char *offs_cnst = NULL;
- char *offs_lea = NULL;
- int scale = 0;
- int isadd = 0;
- int dolea = 0;
- ir_node *left, *right, *temp;
- ir_node *base, *index;
- ia32_am_flavour_t am_flav;
-
- if (is_ia32_Add(irn))
- isadd = 1;
-
- left = get_irn_n(irn, 2);
- right = get_irn_n(irn, 3);
-
- /* "normalize" arguments in case of add with two operands */
- if (isadd && ! be_is_NoReg(babi, right)) {
- /* put LEA == ia32_am_O as right operand */
- if (is_ia32_Lea(left) && get_ia32_am_flavour(left) == ia32_am_O) {
- set_irn_n(irn, 2, right);
- set_irn_n(irn, 3, left);
- temp = left;
- left = right;
- right = temp;
- }
-
- /* put LEA != ia32_am_O as left operand */
- if (is_ia32_Lea(right) && get_ia32_am_flavour(right) != ia32_am_O) {
- set_irn_n(irn, 2, right);
- set_irn_n(irn, 3, left);
- temp = left;
- left = right;
- right = temp;
- }
+ copy_mark(node, xor);
+ be_peephole_exchange(node, xor);
+}
- /* put SHL as left operand iff left is NOT a LEA */
- if (! is_ia32_Lea(left) && pred_is_specific_node(right, is_ia32_Shl)) {
- set_irn_n(irn, 2, right);
- set_irn_n(irn, 3, left);
- temp = left;
- left = right;
- right = temp;
- }
- }
+static INLINE int is_noreg(ia32_code_gen_t *cg, const ir_node *node)
+{
+ return node == cg->noreg_gp;
+}
- base = left;
- index = noreg;
- offs = NULL;
- scale = 0;
- am_flav = 0;
+static ir_node *create_immediate_from_int(int val)
+{
+ ir_graph *irg = current_ir_graph;
+ ir_node *start_block = get_irg_start_block(irg);
+ ir_node *immediate = new_rd_ia32_Immediate(NULL, irg, start_block, NULL,
+ 0, val);
+ arch_set_irn_register(immediate, &ia32_gp_regs[REG_GP_NOREG]);
- /* check if operand is either const */
- if (is_ia32_ImmConst(irn) || is_ia32_ImmSymConst(irn)) {
- DBG((mod, LEVEL_1, "\tfound op with imm"));
+ return immediate;
+}
- offs_cnst = get_ia32_cnst(irn);
- dolea = 1;
- }
+static ir_node *create_immediate_from_am(const ir_node *node)
+{
+ ir_graph *irg = get_irn_irg(node);
+ ir_node *block = get_nodes_block(node);
+ int offset = get_ia32_am_offs_int(node);
+ int sc_sign = is_ia32_am_sc_sign(node);
+ ir_entity *entity = get_ia32_am_sc(node);
+ ir_node *res;
+
+ res = new_rd_ia32_Immediate(NULL, irg, block, entity, sc_sign, offset);
+ arch_set_irn_register(res, &ia32_gp_regs[REG_GP_NOREG]);
+ return res;
+}
- /* determine the operand which needs to be checked */
- if (be_is_NoReg(babi, right)) {
- temp = left;
- }
- else {
- temp = right;
- }
+static int is_am_one(const ir_node *node)
+{
+ int offset = get_ia32_am_offs_int(node);
+ ir_entity *entity = get_ia32_am_sc(node);
- /* check if right operand is AMConst (LEA with ia32_am_O) */
- if (is_ia32_Lea(temp) && get_ia32_am_flavour(temp) == ia32_am_O) {
- DBG((mod, LEVEL_1, "\tgot op with LEA am_O"));
+ return offset == 1 && entity == NULL;
+}
- offs_lea = get_ia32_am_offs(temp);
- dolea = 1;
- }
+static int is_am_minus_one(const ir_node *node)
+{
+ int offset = get_ia32_am_offs_int(node);
+ ir_entity *entity = get_ia32_am_sc(node);
- if (isadd) {
- /* default for add -> make right operand to index */
- index = right;
- dolea = 1;
+ return offset == -1 && entity == NULL;
+}
- DBG((mod, LEVEL_1, "\tgot LEA candidate with index %+F\n", index));
+/**
+ * Transforms a LEA into an Add or SHL if possible.
+ */
+static void peephole_ia32_Lea(ir_node *node)
+{
+ ir_graph *irg = current_ir_graph;
+ ir_node *base;
+ ir_node *index;
+ const arch_register_t *base_reg;
+ const arch_register_t *index_reg;
+ const arch_register_t *out_reg;
+ int scale;
+ int has_immediates;
+ ir_node *op1;
+ ir_node *op2;
+ dbg_info *dbgi;
+ ir_node *block;
+ ir_node *res;
+ ir_node *noreg;
+ ir_node *nomem;
+
+ assert(is_ia32_Lea(node));
+
+ /* we can only do this if are allowed to globber the flags */
+ if(be_peephole_get_value(CLASS_ia32_flags, REG_EFLAGS) != NULL)
+ return;
- /* determine the operand which needs to be checked */
- temp = left;
- if (is_ia32_Lea(left)) {
- temp = right;
- }
+ base = get_irn_n(node, n_ia32_Lea_base);
+ index = get_irn_n(node, n_ia32_Lea_index);
- /* check for SHL 1,2,3 */
- if (pred_is_specific_node(temp, is_ia32_Shl)) {
- temp = get_Proj_pred(temp);
+ if(is_noreg(cg, base)) {
+ base = NULL;
+ base_reg = NULL;
+ } else {
+ base_reg = arch_get_irn_register(base);
+ }
+ if(is_noreg(cg, index)) {
+ index = NULL;
+ index_reg = NULL;
+ } else {
+ index_reg = arch_get_irn_register(index);
+ }
- if (get_ia32_Immop_tarval(temp)) {
- scale = get_tarval_long(get_ia32_Immop_tarval(temp));
+ if(base == NULL && index == NULL) {
+ /* we shouldn't construct these in the first place... */
+#ifdef DEBUG_libfirm
+ ir_fprintf(stderr, "Optimisation warning: found immediate only lea\n");
+#endif
+ return;
+ }
- if (scale <= 3) {
- index = get_irn_n(temp, 2);
+ out_reg = arch_get_irn_register(node);
+ scale = get_ia32_am_scale(node);
+ assert(!is_ia32_need_stackent(node) || get_ia32_frame_ent(node) != NULL);
+ /* check if we have immediates values (frame entities should already be
+ * expressed in the offsets) */
+ if(get_ia32_am_offs_int(node) != 0 || get_ia32_am_sc(node) != NULL) {
+ has_immediates = 1;
+ } else {
+ has_immediates = 0;
+ }
- DBG((mod, LEVEL_1, "\tgot scaled index %+F\n", index));
- }
+ /* we can transform leas where the out register is the same as either the
+ * base or index register back to an Add or Shl */
+ if(out_reg == base_reg) {
+ if(index == NULL) {
+#ifdef DEBUG_libfirm
+ if(!has_immediates) {
+ ir_fprintf(stderr, "Optimisation warning: found lea which is "
+ "just a copy\n");
}
+#endif
+ op1 = base;
+ goto make_add_immediate;
}
-
- /* fix base */
- if (! be_is_NoReg(babi, index)) {
- /* if we have index, but left == right -> no base */
- if (left == right) {
- base = noreg;
- }
- else if (! is_ia32_Lea(left) && (index != right)) {
- /* index != right -> we found a good Shl */
- /* left != LEA -> this Shl was the left operand */
- /* -> base is right operand */
- base = right;
+ if(scale == 0 && !has_immediates) {
+ op1 = base;
+ op2 = index;
+ goto make_add;
+ }
+ /* can't create an add */
+ return;
+ } else if(out_reg == index_reg) {
+ if(base == NULL) {
+ if(has_immediates && scale == 0) {
+ op1 = index;
+ goto make_add_immediate;
+ } else if(!has_immediates && scale > 0) {
+ op1 = index;
+ op2 = create_immediate_from_int(scale);
+ goto make_shl;
+ } else if(!has_immediates) {
+#ifdef DEBUG_libfirm
+ ir_fprintf(stderr, "Optimisation warning: found lea which is "
+ "just a copy\n");
+#endif
}
+ } else if(scale == 0 && !has_immediates) {
+ op1 = index;
+ op2 = base;
+ goto make_add;
}
+ /* can't create an add */
+ return;
+ } else {
+ /* can't create an add */
+ return;
}
- /* Try to assimilate a LEA as left operand */
- if (is_ia32_Lea(left) && (get_ia32_am_flavour(left) != ia32_am_O)) {
- am_flav = get_ia32_am_flavour(left);
-
- /* If we have an Add with a real right operand (not NoReg) and */
- /* the LEA contains already an index calculation then we create */
- /* a new LEA. */
- /* If the LEA contains already a frame_entity then we also */
- /* create a new one otherwise we would loose it. */
- if ((isadd && !be_is_NoReg(babi, index) && (am_flav & ia32_am_I)) ||
- get_ia32_frame_ent(left))
- {
- DBG((mod, LEVEL_1, "\tleave old LEA, creating new one\n"));
+make_add_immediate:
+ if(ia32_cg_config.use_incdec) {
+ if(is_am_one(node)) {
+ dbgi = get_irn_dbg_info(node);
+ block = get_nodes_block(node);
+ res = new_rd_ia32_Inc(dbgi, irg, block, op1);
+ arch_set_irn_register(res, out_reg);
+ goto exchange;
}
- else {
- DBG((mod, LEVEL_1, "\tgot LEA as left operand ... assimilating\n"));
- offs = get_ia32_am_offs(left);
- base = get_irn_n(left, 0);
- index = get_irn_n(left, 1);
- scale = get_ia32_am_scale(left);
+ if(is_am_minus_one(node)) {
+ dbgi = get_irn_dbg_info(node);
+ block = get_nodes_block(node);
+ res = new_rd_ia32_Dec(dbgi, irg, block, op1);
+ arch_set_irn_register(res, out_reg);
+ goto exchange;
}
}
+ op2 = create_immediate_from_am(node);
+
+make_add:
+ dbgi = get_irn_dbg_info(node);
+ block = get_nodes_block(node);
+ noreg = ia32_new_NoReg_gp(cg);
+ nomem = new_NoMem();
+ res = new_rd_ia32_Add(dbgi, irg, block, noreg, noreg, nomem, op1, op2);
+ arch_set_irn_register(res, out_reg);
+ set_ia32_commutative(res);
+ goto exchange;
+
+make_shl:
+ dbgi = get_irn_dbg_info(node);
+ block = get_nodes_block(node);
+ noreg = ia32_new_NoReg_gp(cg);
+ nomem = new_NoMem();
+ res = new_rd_ia32_Shl(dbgi, irg, block, op1, op2);
+ arch_set_irn_register(res, out_reg);
+ goto exchange;
+
+exchange:
+ SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(cg, node));
+
+ /* add new ADD/SHL to schedule */
+ DBG_OPT_LEA2ADD(node, res);
+
+ /* exchange the Add and the LEA */
+ sched_add_before(node, res);
+ copy_mark(node, res);
+ be_peephole_exchange(node, res);
+}
- /* ok, we can create a new LEA */
- if (dolea) {
- res = new_rd_ia32_Lea(dbg, irg, block, base, index, mode_Is);
-
- /* add the old offset of a previous LEA */
- if (offs) {
- add_ia32_am_offs(res, offs);
- }
+/**
+ * Split a Imul mem, imm into a Load mem and Imul reg, imm if possible.
+ */
+static void peephole_ia32_Imul_split(ir_node *imul)
+{
+ const ir_node *right = get_irn_n(imul, n_ia32_IMul_right);
+ const arch_register_t *reg;
+ ir_node *res;
- /* add the new offset */
- if (isadd) {
- if (offs_cnst) {
- add_ia32_am_offs(res, offs_cnst);
- }
- if (offs_lea) {
- add_ia32_am_offs(res, offs_lea);
- }
- }
- else {
- /* either lea_O-cnst, -cnst or -lea_O */
- if (offs_cnst) {
- if (offs_lea) {
- add_ia32_am_offs(res, offs_lea);
- }
+ if (!is_ia32_Immediate(right) || get_ia32_op_type(imul) != ia32_AddrModeS) {
+ /* no memory, imm form ignore */
+ return;
+ }
+ /* we need a free register */
+ reg = get_free_gp_reg();
+ if (reg == NULL)
+ return;
- sub_ia32_am_offs(res, offs_cnst);
- }
- else {
- sub_ia32_am_offs(res, offs_lea);
- }
- }
+ /* fine, we can rebuild it */
+ res = turn_back_am(imul);
+ arch_set_irn_register(res, reg);
+}
- /* copy the frame entity (could be set in case of Add */
- /* which was a FrameAddr) */
- set_ia32_frame_ent(res, get_ia32_frame_ent(irn));
+/**
+ * Replace xorps r,r and xorpd r,r by pxor r,r
+ */
+static void peephole_ia32_xZero(ir_node *xor) {
+ set_irn_op(xor, op_ia32_xPzero);
+}
- if (is_ia32_use_frame(irn))
- set_ia32_use_frame(res);
+/**
+ * Register a peephole optimisation function.
+ */
+static void register_peephole_optimisation(ir_op *op, peephole_opt_func func) {
+ assert(op->ops.generic == NULL);
+ op->ops.generic = (op_func)func;
+}
- /* set scale */
- set_ia32_am_scale(res, scale);
+/* Perform peephole-optimizations. */
+void ia32_peephole_optimization(ia32_code_gen_t *new_cg)
+{
+ cg = new_cg;
+
+ /* register peephole optimisations */
+ clear_irp_opcodes_generic_func();
+ register_peephole_optimisation(op_ia32_Const, peephole_ia32_Const);
+ register_peephole_optimisation(op_be_IncSP, peephole_be_IncSP);
+ register_peephole_optimisation(op_ia32_Lea, peephole_ia32_Lea);
+ register_peephole_optimisation(op_ia32_Cmp, peephole_ia32_Cmp);
+ register_peephole_optimisation(op_ia32_Cmp8Bit, peephole_ia32_Cmp);
+ register_peephole_optimisation(op_ia32_Test, peephole_ia32_Test);
+ register_peephole_optimisation(op_ia32_Test8Bit, peephole_ia32_Test);
+ register_peephole_optimisation(op_be_Return, peephole_ia32_Return);
+ if (! ia32_cg_config.use_imul_mem_imm32)
+ register_peephole_optimisation(op_ia32_IMul, peephole_ia32_Imul_split);
+ if (ia32_cg_config.use_pxor)
+ register_peephole_optimisation(op_ia32_xZero, peephole_ia32_xZero);
+
+ be_peephole_opt(cg->birg);
+}
- am_flav = ia32_am_N;
- /* determine new am flavour */
- if (offs || offs_cnst || offs_lea) {
- am_flav |= ia32_O;
- }
- if (! be_is_NoReg(babi, base)) {
- am_flav |= ia32_B;
- }
- if (! be_is_NoReg(babi, index)) {
- am_flav |= ia32_I;
- }
- if (scale > 0) {
- am_flav |= ia32_S;
+/**
+ * Removes node from schedule if it is not used anymore. If irn is a mode_T node
+ * all it's Projs are removed as well.
+ * @param irn The irn to be removed from schedule
+ */
+static INLINE void try_kill(ir_node *node)
+{
+ if(get_irn_mode(node) == mode_T) {
+ const ir_edge_t *edge, *next;
+ foreach_out_edge_safe(node, edge, next) {
+ ir_node *proj = get_edge_src_irn(edge);
+ try_kill(proj);
}
- set_ia32_am_flavour(res, am_flav);
+ }
- set_ia32_op_type(res, ia32_AddrModeS);
+ if(get_irn_n_edges(node) != 0)
+ return;
- DBG((mod, LEVEL_1, "\tLEA [%+F + %+F * %d + %s]\n", base, index, scale, get_ia32_am_offs(res)));
+ if (sched_is_scheduled(node)) {
+ sched_remove(node);
+ }
- /* get the result Proj of the Add/Sub */
- irn = get_res_proj(irn);
+ kill_node(node);
+}
- assert(irn && "Couldn't find result proj");
+static void optimize_conv_store(ir_node *node)
+{
+ ir_node *pred;
+ ir_node *pred_proj;
+ ir_mode *conv_mode;
+ ir_mode *store_mode;
- /* exchange the old op with the new LEA */
- exchange(irn, res);
+ if(!is_ia32_Store(node) && !is_ia32_Store8Bit(node))
+ return;
+
+ assert(n_ia32_Store_val == n_ia32_Store8Bit_val);
+ pred_proj = get_irn_n(node, n_ia32_Store_val);
+ if(is_Proj(pred_proj)) {
+ pred = get_Proj_pred(pred_proj);
+ } else {
+ pred = pred_proj;
}
+ if(!is_ia32_Conv_I2I(pred) && !is_ia32_Conv_I2I8Bit(pred))
+ return;
+ if(get_ia32_op_type(pred) != ia32_Normal)
+ return;
- return res;
+ /* the store only stores the lower bits, so we only need the conv
+ * it it shrinks the mode */
+ conv_mode = get_ia32_ls_mode(pred);
+ store_mode = get_ia32_ls_mode(node);
+ if(get_mode_size_bits(conv_mode) < get_mode_size_bits(store_mode))
+ return;
+
+ set_irn_n(node, n_ia32_Store_val, get_irn_n(pred, n_ia32_Conv_I2I_val));
+ if(get_irn_n_edges(pred_proj) == 0) {
+ kill_node(pred_proj);
+ if(pred != pred_proj)
+ kill_node(pred);
+ }
}
-/**
- * Optimizes a pattern around irn to address mode if possible.
- */
-void ia32_optimize_am(ir_node *irn, void *env) {
- ia32_code_gen_t *cg = env;
- firm_dbg_module_t *mod = cg->mod;
- ir_node *res = irn;
- be_abi_irg_t *babi = cg->birg->abi;
- dbg_info *dbg;
- ir_mode *mode;
- ir_node *block, *noreg_gp, *noreg_fp;
- ir_node *left, *right, *temp;
- ir_node *store, *load, *mem_proj;
- ir_node *succ, *addr_b, *addr_i;
- int check_am_src = 0;
-
- if (! is_ia32_irn(irn))
+static void optimize_load_conv(ir_node *node)
+{
+ ir_node *pred, *predpred;
+ ir_mode *load_mode;
+ ir_mode *conv_mode;
+
+ if (!is_ia32_Conv_I2I(node) && !is_ia32_Conv_I2I8Bit(node))
return;
- dbg = get_irn_dbg_info(irn);
- mode = get_irn_mode(irn);
- block = get_nodes_block(irn);
- noreg_gp = ia32_new_NoReg_gp(cg);
- noreg_fp = ia32_new_NoReg_fp(cg);
-
- DBG((mod, LEVEL_1, "checking for AM\n"));
-
- /* 1st part: check for address calculations and transform the into Lea */
-
- /* Following cases can occur: */
- /* - Sub (l, imm) -> LEA [base - offset] */
- /* - Sub (l, r == LEA with ia32_am_O) -> LEA [base - offset] */
- /* - Add (l, imm) -> LEA [base + offset] */
- /* - Add (l, r == LEA with ia32_am_O) -> LEA [base + offset] */
- /* - Add (l == LEA with ia32_am_O, r) -> LEA [base + offset] */
- /* - Add (l, r) -> LEA [base + index * scale] */
- /* with scale > 1 iff l/r == shl (1,2,3) */
-
- if (is_ia32_Sub(irn) || is_ia32_Add(irn)) {
- left = get_irn_n(irn, 2);
- right = get_irn_n(irn, 3);
-
- /* Do not try to create a LEA if one of the operands is a Load. */
- /* check is irn is a candidate for address calculation */
- if (is_candidate(block, irn, 1)) {
- DBG((mod, LEVEL_1, "\tfound address calculation candidate %+F ... ", irn));
- res = fold_addr(babi, irn, mod, noreg_gp);
-
- if (res == irn)
- DB((mod, LEVEL_1, "transformed into %+F\n", res));
- else
- DB((mod, LEVEL_1, "not transformed\n"));
- }
- }
+ assert(n_ia32_Conv_I2I_val == n_ia32_Conv_I2I8Bit_val);
+ pred = get_irn_n(node, n_ia32_Conv_I2I_val);
+ if(!is_Proj(pred))
+ return;
- /* 2nd part: fold following patterns: */
- /* - Load -> LEA into Load } TODO: If the LEA is used by more than one Load/Store */
- /* - Store -> LEA into Store } it might be better to keep the LEA */
- /* - op -> Load into AMop with am_Source */
- /* conditions: */
- /* - op is am_Source capable AND */
- /* - the Load is only used by this op AND */
- /* - the Load is in the same block */
- /* - Store -> op -> Load into AMop with am_Dest */
- /* conditions: */
- /* - op is am_Dest capable AND */
- /* - the Store uses the same address as the Load AND */
- /* - the Load is only used by this op AND */
- /* - the Load and Store are in the same block AND */
- /* - nobody else uses the result of the op */
-
- if ((res == irn) && (get_ia32_am_support(irn) != ia32_am_None) && !is_ia32_Lea(irn)) {
- /* 1st: check for Load/Store -> LEA */
- if (is_ia32_Ld(irn) || is_ia32_St(irn) || is_ia32_Store8Bit(irn)) {
- left = get_irn_n(irn, 0);
-
- if (is_ia32_Lea(left)) {
- DBG((mod, LEVEL_1, "\nmerging %+F into %+F\n", left, irn));
-
- /* get the AM attributes from the LEA */
- add_ia32_am_offs(irn, get_ia32_am_offs(left));
- set_ia32_am_scale(irn, get_ia32_am_scale(left));
- set_ia32_am_flavour(irn, get_ia32_am_flavour(left));
-
- set_ia32_op_type(irn, is_ia32_Ld(irn) ? ia32_AddrModeS : ia32_AddrModeD);
-
- /* set base and index */
- set_irn_n(irn, 0, get_irn_n(left, 0));
- set_irn_n(irn, 1, get_irn_n(left, 1));
-
- /* clear remat flag */
- set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable);
+ predpred = get_Proj_pred(pred);
+ if(!is_ia32_Load(predpred))
+ return;
+
+ /* the load is sign extending the upper bits, so we only need the conv
+ * if it shrinks the mode */
+ load_mode = get_ia32_ls_mode(predpred);
+ conv_mode = get_ia32_ls_mode(node);
+ if(get_mode_size_bits(conv_mode) < get_mode_size_bits(load_mode))
+ return;
+
+ if(get_mode_sign(conv_mode) != get_mode_sign(load_mode)) {
+ /* change the load if it has only 1 user */
+ if(get_irn_n_edges(pred) == 1) {
+ ir_mode *newmode;
+ if(get_mode_sign(conv_mode)) {
+ newmode = find_signed_mode(load_mode);
+ } else {
+ newmode = find_unsigned_mode(load_mode);
}
+ assert(newmode != NULL);
+ set_ia32_ls_mode(predpred, newmode);
+ } else {
+ /* otherwise we have to keep the conv */
+ return;
}
- /* check if the node is an address mode candidate */
- else if (is_candidate(block, irn, 0)) {
- DBG((mod, LEVEL_1, "\tfound address mode candidate %+F ... ", irn));
-
- left = get_irn_n(irn, 2);
- if (get_irn_arity(irn) == 4) {
- /* it's an "unary" operation */
- right = left;
- }
- else {
- right = get_irn_n(irn, 3);
- }
+ }
- /* normalize commutative ops */
- if (node_is_comm(irn)) {
- /* Assure that right operand is always a Load if there is one */
- /* because non-commutative ops can only use Dest AM if the right */
- /* operand is a load, so we only need to check right operand. */
- if (pred_is_specific_nodeblock(block, left, is_ia32_Ld))
- {
- set_irn_n(irn, 2, right);
- set_irn_n(irn, 3, left);
-
- temp = left;
- left = right;
- right = temp;
- }
- }
+ /* kill the conv */
+ exchange(node, pred);
+}
- /* check for Store -> op -> Load */
-
- /* Store -> op -> Load optimization is only possible if supported by op */
- /* and if right operand is a Load */
- if ((get_ia32_am_support(irn) & ia32_am_Dest) &&
- pred_is_specific_nodeblock(block, right, is_ia32_Ld))
- {
-
- /* An address mode capable op always has a result Proj. */
- /* If this Proj is used by more than one other node, we don't need to */
- /* check further, otherwise we check for Store and remember the address, */
- /* the Store points to. */
-
- succ = get_res_proj(irn);
- assert(succ && "Couldn't find result proj");
-
- addr_b = NULL;
- addr_i = NULL;
- store = NULL;
-
- /* now check for users and Store */
- if (ia32_get_irn_n_edges(succ) == 1) {
- succ = get_edge_src_irn(get_irn_out_edge_first(succ));
-
- if (is_ia32_fStore(succ) || is_ia32_Store(succ)) {
- store = succ;
- addr_b = get_irn_n(store, 0);
-
- /* Could be that the Store is connected to the address */
- /* calculating LEA while the Load is already transformed. */
- if (is_ia32_Lea(addr_b)) {
- succ = addr_b;
- addr_b = get_irn_n(succ, 0);
- addr_i = get_irn_n(succ, 1);
- }
- else {
- addr_i = noreg_gp;
- }
- }
- }
+static void optimize_conv_conv(ir_node *node)
+{
+ ir_node *pred_proj, *pred, *result_conv;
+ ir_mode *pred_mode, *conv_mode;
+ int conv_mode_bits;
+ int pred_mode_bits;
- if (store) {
- /* we found a Store as single user: Now check for Load */
-
- /* Extra check for commutative ops with two Loads */
- /* -> put the interesting Load right */
- if (node_is_comm(irn) &&
- pred_is_specific_nodeblock(block, left, is_ia32_Ld))
- {
- if ((addr_b == get_irn_n(get_Proj_pred(left), 0)) &&
- (addr_i == get_irn_n(get_Proj_pred(left), 1)))
- {
- /* We exchange left and right, so it's easier to kill */
- /* the correct Load later and to handle unary operations. */
- set_irn_n(irn, 2, right);
- set_irn_n(irn, 3, left);
-
- temp = left;
- left = right;
- right = temp;
- }
- }
-
- /* skip the Proj for easier access */
- load = get_Proj_pred(right);
-
- /* Compare Load and Store address */
- if (load_store_addr_is_equal(load, store, addr_b, addr_i)) {
- /* Right Load is from same address, so we can */
- /* disconnect the Load and Store here */
-
- /* set new base, index and attributes */
- set_irn_n(irn, 0, addr_b);
- set_irn_n(irn, 1, addr_i);
- add_ia32_am_offs(irn, get_ia32_am_offs(load));
- set_ia32_am_scale(irn, get_ia32_am_scale(load));
- set_ia32_am_flavour(irn, get_ia32_am_flavour(load));
- set_ia32_op_type(irn, ia32_AddrModeD);
- set_ia32_frame_ent(irn, get_ia32_frame_ent(load));
- set_ia32_ls_mode(irn, get_ia32_ls_mode(load));
-
- if (is_ia32_use_frame(load))
- set_ia32_use_frame(irn);
-
- /* connect to Load memory and disconnect Load */
- if (get_irn_arity(irn) == 5) {
- /* binary AMop */
- set_irn_n(irn, 4, get_irn_n(load, 2));
- set_irn_n(irn, 3, noreg_gp);
- }
- else {
- /* unary AMop */
- set_irn_n(irn, 3, get_irn_n(load, 2));
- set_irn_n(irn, 2, noreg_gp);
- }
-
- /* connect the memory Proj of the Store to the op */
- mem_proj = get_mem_proj(store);
- set_Proj_pred(mem_proj, irn);
- set_Proj_proj(mem_proj, 1);
-
- /* clear remat flag */
- set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable);
-
- DB((mod, LEVEL_1, "merged with %+F and %+F into dest AM\n", load, store));
- }
- } /* if (store) */
- else if (get_ia32_am_support(irn) & ia32_am_Source) {
- /* There was no store, check if we still can optimize for source address mode */
- check_am_src = 1;
- }
- } /* if (support AM Dest) */
- else if (get_ia32_am_support(irn) & ia32_am_Source) {
- /* op doesn't support am AM Dest -> check for AM Source */
- check_am_src = 1;
- }
+ if (!is_ia32_Conv_I2I(node) && !is_ia32_Conv_I2I8Bit(node))
+ return;
- /* normalize commutative ops */
- if (node_is_comm(irn)) {
- /* Assure that left operand is always a Load if there is one */
- /* because non-commutative ops can only use Source AM if the */
- /* left operand is a Load, so we only need to check the left */
- /* operand afterwards. */
- if (pred_is_specific_nodeblock(block, right, is_ia32_Ld)) {
- set_irn_n(irn, 2, right);
- set_irn_n(irn, 3, left);
-
- temp = left;
- left = right;
- right = temp;
- }
- }
+ assert(n_ia32_Conv_I2I_val == n_ia32_Conv_I2I8Bit_val);
+ pred_proj = get_irn_n(node, n_ia32_Conv_I2I_val);
+ if(is_Proj(pred_proj))
+ pred = get_Proj_pred(pred_proj);
+ else
+ pred = pred_proj;
- /* optimize op -> Load iff Load is only used by this op */
- /* and left operand is a Load which only used by this irn */
- if (check_am_src &&
- pred_is_specific_nodeblock(block, left, is_ia32_Ld) &&
- (ia32_get_irn_n_edges(left) == 1))
- {
- left = get_Proj_pred(left);
-
- addr_b = get_irn_n(left, 0);
- addr_i = get_irn_n(left, 1);
-
- /* set new base, index and attributes */
- set_irn_n(irn, 0, addr_b);
- set_irn_n(irn, 1, addr_i);
- add_ia32_am_offs(irn, get_ia32_am_offs(left));
- set_ia32_am_scale(irn, get_ia32_am_scale(left));
- set_ia32_am_flavour(irn, get_ia32_am_flavour(left));
- set_ia32_op_type(irn, ia32_AddrModeS);
- set_ia32_frame_ent(irn, get_ia32_frame_ent(left));
- set_ia32_ls_mode(irn, get_ia32_ls_mode(left));
-
- /* clear remat flag */
- set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable);
-
- if (is_ia32_use_frame(left))
- set_ia32_use_frame(irn);
-
- /* connect to Load memory */
- if (get_irn_arity(irn) == 5) {
- /* binary AMop */
- set_irn_n(irn, 4, get_irn_n(left, 2));
- }
- else {
- /* unary AMop */
- set_irn_n(irn, 3, get_irn_n(left, 2));
- }
+ if(!is_ia32_Conv_I2I(pred) && !is_ia32_Conv_I2I8Bit(pred))
+ return;
- /* disconnect from Load */
- set_irn_n(irn, 2, noreg_gp);
+ /* we know that after a conv, the upper bits are sign extended
+ * so we only need the 2nd conv if it shrinks the mode */
+ conv_mode = get_ia32_ls_mode(node);
+ conv_mode_bits = get_mode_size_bits(conv_mode);
+ pred_mode = get_ia32_ls_mode(pred);
+ pred_mode_bits = get_mode_size_bits(pred_mode);
+
+ if(conv_mode_bits == pred_mode_bits
+ && get_mode_sign(conv_mode) == get_mode_sign(pred_mode)) {
+ result_conv = pred_proj;
+ } else if(conv_mode_bits <= pred_mode_bits) {
+ /* if 2nd conv is smaller then first conv, then we can always take the
+ * 2nd conv */
+ if(get_irn_n_edges(pred_proj) == 1) {
+ result_conv = pred_proj;
+ set_ia32_ls_mode(pred, conv_mode);
+
+ /* Argh:We must change the opcode to 8bit AND copy the register constraints */
+ if (get_mode_size_bits(conv_mode) == 8) {
+ set_irn_op(pred, op_ia32_Conv_I2I8Bit);
+ set_ia32_in_req_all(pred, get_ia32_in_req_all(node));
+ }
+ } else {
+ /* we don't want to end up with 2 loads, so we better do nothing */
+ if(get_irn_mode(pred) == mode_T) {
+ return;
+ }
- /* If Load has a memory Proj, connect it to the op */
- mem_proj = get_mem_proj(left);
- if (mem_proj) {
- set_Proj_pred(mem_proj, irn);
- set_Proj_proj(mem_proj, 1);
- }
+ result_conv = exact_copy(pred);
+ set_ia32_ls_mode(result_conv, conv_mode);
- DB((mod, LEVEL_1, "merged with %+F into source AM\n", left));
+ /* Argh:We must change the opcode to 8bit AND copy the register constraints */
+ if (get_mode_size_bits(conv_mode) == 8) {
+ set_irn_op(result_conv, op_ia32_Conv_I2I8Bit);
+ set_ia32_in_req_all(result_conv, get_ia32_in_req_all(node));
}
}
+ } else {
+ /* if both convs have the same sign, then we can take the smaller one */
+ if(get_mode_sign(conv_mode) == get_mode_sign(pred_mode)) {
+ result_conv = pred_proj;
+ } else {
+ /* no optimisation possible if smaller conv is sign-extend */
+ if(mode_is_signed(pred_mode)) {
+ return;
+ }
+ /* we can take the smaller conv if it is unsigned */
+ result_conv = pred_proj;
+ }
}
+
+ /* kill the conv */
+ exchange(node, result_conv);
+
+ if(get_irn_n_edges(pred_proj) == 0) {
+ kill_node(pred_proj);
+ if(pred != pred_proj)
+ kill_node(pred);
+ }
+ optimize_conv_conv(result_conv);
+}
+
+static void optimize_node(ir_node *node, void *env)
+{
+ (void) env;
+
+ optimize_load_conv(node);
+ optimize_conv_store(node);
+ optimize_conv_conv(node);
+}
+
+/**
+ * Performs conv and address mode optimization.
+ */
+void ia32_optimize_graph(ia32_code_gen_t *cg)
+{
+ irg_walk_blkwise_graph(cg->irg, NULL, optimize_node, cg);
+
+ if (cg->dump)
+ be_dump(cg->irg, "-opt", dump_ir_block_graph_sched);
+}
+
+void ia32_init_optimize(void)
+{
+ FIRM_DBG_REGISTER(dbg, "firm.be.ia32.optimize");
}
projects / libfirm / blobdiff